ReactVAU: A Slow-Fast Decoupled Framework for Streaming Video Anomaly Understanding
ReactVAU is built to catch and explain anomalies in live video streams without running a heavy multimodal model on every frame.
The paper says current video anomaly systems often depend on offline, non-causal sampling, making them a poor fit for live surveillance. ReactVAU splits the job into a lightweight fast detector, persistent anomaly-aware memory, and a slow reasoning module that activates only when something looks suspicious. The authors report competitive anomaly detection and causal reasoning across multiple benchmarks, with better efficiency from fewer heavyweight model calls. HF Daily Papers' note
The paper says current video anomaly systems often depend on offline, non-causal sampling, making them a poor fit for live surveillance. ReactVAU splits the job into a lightweight fast detector, persistent anomaly-aware memory, and a slow reasoning module that activates only when something looks suspicious. The authors report competitive anomaly detection and causal reasoning across multiple benchmarks, with better efficiency from fewer heavyweight model calls. HF Daily Papers' note
score 4